Scripts for predicting how short fragments of natural proteins bind to full-length proteins, as described in the manuscript. This program is built on top of MMseqs2 and ColabFold, extending them to efficiently predict interactions between a full-length protein and fragments derived from a protein.
This code is associated with the following article: A. Savinov, S. Swanson, A. E. Keating, G.-W. Li. High-throughput computational discovery of inhibitory protein fragments with AlphaFold. bioRxiv (2023). doi: 10.1101/2023.12.19.572389. https://www.biorxiv.org/content/10.1101/2023.12.19.572389v2
Please cite this article if you make use of FragFold.
Associated Source Data can also be found here:
https://figshare.com/articles/dataset/Source_Data_for_Savinov_and_Swanson_et_al_2023/24841269
doi: 10.6084/m9.figshare.24841269
The locally installable version of ColabFold is used to predict structures. Go to the repo and follow the instructions to install ColabFold on your system (this can take >1 hr). You can ignore the recommendation to edit your ~/.bashrc to add local ColabFold to your PATH, this is handled by FragFold.
We developed and tested FragFold on this specific commit. Note that CUDA 11.8 is required to use a GPU with ColabFold.
Log the output while installing to verify that each step of the process completed successfully, this can be particularly helpful for debugging issues later on.
bash install_colabbatch_linux.sh | tee install_colabbatch_linux.logOnce installation has completed successfully, run a small test job to verify that it's working. The following example is taken from the localcolabfold README.
/path/to/installation/colabfold_batch example.fasta colabfold_batch_testWhere example.fasta is a short amino acid sequence
>sp|P61823
MALKSLVLLSLLVLVLLLVRVQPSLGKETAAAKFERQHMDSSTSAASSSNYCNQMMKSRN
LTKDRCKPVNTFVHESLADVQAVCSQKNVACKNGQTNCYQSYSTMSITDCRETGSSKYPN
CAYKTTQANKHIIVACEGNPYVPVHFDASVIf you're running on a system with a NVIDIA gpu, look for Running on GPU in the output. If you see that not GPU is detected, check if one is currently available with nvidia-smi. If you're still having issues, look back at the logs to check if the version of tensorflow that was installed is compatible with CUDA 11.8.
Use the following commands to clone the repo and install FragFold and its dependencies on your computer.
git clone https://github.com/swanss/FragFold.git
cd FragFold
bash install_fragfold.shinstall_fragfold.sh uses conda to set up an environment containing FragFold as well as nextflow, the workflow system used to run jobs. The script should print installation complete, if not, there was an issue during one of steps.
FragFold uses nextflow to manage the execution of the pipeline. MSA building with MMseq2s, MSA processing/concatenation, ColabFold, and the analysis and aggregation are each defined as individual processes in FragFold/nextflow/modules.nf and then used to construct workflows for modeling homomeric or heteromeric interactions between fragments and full-length proteins, e.g. FragFold/nextflow/main.nf. The details of execution, such as whether a process should be run locally or submitted to a job queue in an HPC environment is controlled by the config file FragFold/nextflow/nextflow.config. The job-specific parameters are defined in a yaml file, for example, FragFold/nextflow/params/ftsZ_monomeric_example.yaml.
The following example demonstrates how to model homomeric interactions between full-length FtsZ (with slight truncations of the terminal residues) with 30aa fragments derived from the protein.
Before running the pipeline, edit FragFold/nextflow/nextflow.config. First, set general parameters according to your install.
// Define system-specific variables for all processes
executor.queueSize = 50
executor.conda = '/home/gridsan/sswanson/mambaforge/envs/fragfold'
env.repo_dir = '/data1/groups/keatinglab/swans/savinovCollaboration/FragFold'
env.colabfold_dir = '/home/gridsan/sswanson/localcolabfold/colabfold-conda'
env.alphafold_params_dir = '/home/gridsan/sswanson/localcolabfold/colabfold'
// Parameters that general don't change (but you can adjust to experiment)
env.model_type = 'alphafold2_ptm' //alt: alphafold2_multimer_v1 alphafold2_multimer_v2 alphafold2_multimer_v3
env.pair_mode = 'unpaired' //alt: paired
// Parameters that govern how the output is stored
env.colabfold_outdir = "output_colabfold"
env.peakprediction_outdir = "output_peakprediction"Tips
conda info --envscolabfold_dir will be determined based on where you ran the installalphafold_params_dir can be found with the colabfold directoryNext you will need to edit the process profiles (in FragFold/nextflow/nextflow.config) according to your HPC environment.
withLabel: cpu_network {
executor = 'slurm'
time = '1h'
cpus = 1
memory = '4 GB'
queue = 'download' // edit this value
clusterOptions = '--ntasks-per-node=1'
}In this example, we define a process profile with the label cpu_network. If you're using a HPC with SLURM, you will only need to change queue to a partition on your HPC (see all partitions with sinfo). Note that the nodes in the partition set for cpu_network must have network access, as this is required for contacting the mmseqs server.
Repeat this for all of the profiles that are defined in the config file. Note that if you have nodes with different memory/cpu/time limits, you will need to adjust these values to match those.
Tips
FragFold/nextflow/modules.nf and replace the label for each process with 'standard'Now we set the job-specific parameters that control FragFold execution. In order to run the example, use FragFold/nextflow/params/ftsZ_monomeric_example.yaml, you will only need to edit the path of the query sequence and experimental to match your local installation.
job_name: ftsZ_test
heteromeric_mode: false
protein_query_seq: /home/gridsan/sswanson/keatinglab_shared/swans/savinovCollaboration/FragFold/input/gene_fasta/ftsZ_A0A140NFM6.fasta
protein_name: ftsZ-coding-EcoliBL21DE3 #"" if you want to use the protein_query_seq filename
fragment_ntermres_start: 260
fragment_ntermres_final: 265
fragment_length: 30
protein_nterm_res: 10
protein_cterm_res: 316
protein_copies: 1
experimental_data: /home/gridsan/sswanson/keatinglab_shared/swans/savinovCollaboration/FragFold/input/inhibitory_data/Savinov_2022_inhib_peptide_mapping.csv
n_contacts: 3
n_weighted_contacts: 3
iptm: 0.3
contact_distance_cutoff: 3.5
contact_distance_cluster_cutoff: 0.6
cluster_peaks_frac_overlap: 0.7Note: if you don't have any experimental data corresponding to the fragments, set the path to nextflow/assets/NO_FILE.
experimental_data: /data1/groups/keatinglab/swans/savinovCollaboration/FragFold/nextflow/assets/NO_FILERun nextflow with the following command:
NEXTFLOWDIR=/home/gridsan/user/FragFold/nextflow #directory containing nextflow scripts
WORKDIR=/home/gridsan/user/FragFold/example #directory where results will be stored
NF_CFG=${NEXTFLOWDIR}/nextflow.config
PARAMS=${NEXTFLOWDIR}/params/ftsZ_monomeric_example.yaml
nextflow run ${NEXTFLOWDIR}/main.nf -w $WORKDIR -c $NF_CFG -params-file $PARAMS -resumeAssuming you have access to GPUs, this should take ~half an hour. After the job is completed you should see output like this:
(fragfold) user@d-19-1-1:/state/partition1/user/user$ nextflow run ${NEXTFLOWDIR}/main.nf -w $WORKDIR -c $NF_CFG -params-file $PARAMS -resume
N E X T F L O W ~ version 23.10.1
Launching `/home/gridsan/user/keatinglab_shared/user/savinovCollaboration/FragFold/nextflow/ftsZ_homomeric_example.nf` [tiny_agnesi] DSL2 - revision: f2c61140f9
executor > slurm (1)
[b5/8446c6] process > build_msa [100%] 1 of 1 ✔
[51/8c5e4b] process > process_msa [100%] 1 of 1 ✔
[66/32bf2a] process > colabfold (5) [100%] 11 of 11 ✔
[89/a44267] process > create_summary_csv [100%] 1 of 1 ✔
[d8/d7g0w3] process > predict_peaks [100%] 1 of 1 ✔To inspect the output of intermediate steps, use the names of the processes to find the directory. For example to find output from colabfold, go to /home/gridsan/user/FragFold/nextflow/practice/66/32bf2a*.
The output csvs (colabfold output: *_results_expmerge.csv, predicted peaks: *_predictalphafoldpeaks_*.csv) are copied to the working directory where the nextflow job was submitted. Note that there is more output contained in the directories, including plots of the weighted contacts by fragment position.
Nextflow uses file locking to store task metadata and can only be run in a directory that has locking. This is not compatible with certain filesystem types, such as Lustre. To check what file systems are available on your system, use findmnt, the output will report the file system type for available directories and also describe whether locking is available in the OPTIONS column (If you see noflock, locking is not supported).
One workaround for this is to run nextflow on a file system that supports locking (e.g. a local filesystem) for task metadata storage while running the processes/and storing the output on a shared lustre directory. This works because the directory where results are stored does not need file locking (as outputs are always stored in separate directories to avoid collisions). To do this, simply run the nextflow command in a directory with locking and add the -w argument to specify the working directory.
As an example, this is how we start nextflow on our HPC:
# request an interactive node
LLsub -i --time=1-00:00 --partition=xeon-p8
# create a new directory on the local filesystem
USER=$(whoami) && mkdir -p /state/partition1/user/$USER && cd /state/partition1/user/$USER && conda activate fragfold
NEXTFLOWDIR=/home/gridsan/user/FragFold/nextflow #directory containing nextflow scripts
WORKDIR=/home/gridsan/user/FragFold/example #directory where results will be stored
NF_CFG=${NEXTFLOWDIR}/nextflow.config
PARAMS=${NEXTFLOWDIR}/params/ftsZ_monomeric_example.yaml
nextflow run ${NEXTFLOWDIR}/main.nf -w $WORKDIR -c $NF_CFG -params-file $PARAMS -resumeFor most users, it will take days for all the submitted colabfold jobs to complete. In order to keep nextflow running until all the processes are complete, we submit it as a job with a long time limit. For help with submitting jobs, see the example script: FragFold/scripts/submit/submit_nextflow.sh. This script supports -resume by copying the task metadata files. To get it running on your HPC, you will need to edit the slurm directives and paths.